SQS: A Multi-Dimensional Framework for Evaluating Software Specifications in AI-Assisted Development
Abhavya Anekverna · Aug 15, 2026
Abstract If AI coding agents need to make a lot of inferences, they are more likely to create incorrect code, but if they have less work to do, theoretically it should be more accurate. In an AI coding agent with little to no human interaction, a specification that doesn't specify a behavior can be interpreted as a decision on implementation, with the agent continuing to work without asking for clarification. If a decision on such a behavior, constraint, interface, or acceptance condition is made, it is called a specification-induced assumption, and it is therefore an implementation decision, distinct from an ordinary implementation decision the specification leaves open (e.g., what front-end framework to use). We hypothesize that the quality of a specification impacts the correctness of implementation in the downstream, as follows: Lower quality specification = fewer assumptions an agent may have to make = less deviation from what the specification actually specifies. Current requirements-quality research has been geared towards human-centred requirements processes, and the studies of ambiguity, completeness and consistency have concentrated on them separately. This paper introduces the Specification Quality Score (SQS), a rubric-anchored, operational measure based on five dimensions of specification quality (Completeness, Unambiguity, Internal Consistency, Intent Traceability, Machine-Actionability) with clear criteria for each band of each dimension. We provide worked examples of the framework and a clear explanation of the meaning of the total score. This is a conceptual paper – it is not an empirical validation study, and does not imply that SQS predicts code generation results. This correlational statement, and the controlled experiment necessary to test it, is explicitly suggested as the next step. Keywords: requirements engineering, specification quality, large language models, AI coding agents, specification-driven development